Selection of Optimal Lag Length in

Cointegrated VAR Models with Weak

Form of Common Cyclical Features
+
Carlos Enrique Carrasco Gutierrez
,
Reinaldo Castro Souza
)
Osmani Teixeira de Carvalho Guilln

Abstract
An important aspect of empirical research based on the vector autoregres-
sive (VAR) model is the choice of the lag order, since all inferences in this
model depend on the correct model specication. There have been many
studies of how to select the lag order of a nonstationary VAR model subject
to cointegration restrictions. In this work, we consider in the model an addi-
tional weak form (WF) restriction of common cyclical features to analyze the
appropriate way to select the correct lag order. We use two methodologies:
the traditional information criteria (AIC, HQ and SC) and an alternative cri-
terion (IC(p; s)) that selects the lag order p and the rank structure s due to
the WF restriction. We use a Monte-Carlo simulation in the analysis. The
results indicate that the cost of ignoring additional WF restrictions in vector
autoregressive modeling can be high, especially when the SC criterion is used.
Keywords: Cointegration; Common Cyclical Features; Reduced Rank Model;
Information Criteria.
JEL Codes: C32, C53.

We are grateful for the comments and suggestions by Joo Victor Issler, Wagner Gaglianone,
Ricardo Cavalcanti, Luiz Renato Lima and many seminar participants at the 2006 Brazilian Econo-
metric Meeting and 2007 European Econometric Meeting in Hungary. We extend special thanks
to Alain Hecq for solving doubts and providing helpful comments. We are responsible for any
remaining errors in this paper. Carlos Enrique also acknowledges nancial support from Getulio
Vargas Foundation and the Oce to Improve University Research (CAPES).
y
FUCAPE Business School, Av. Fernando Ferrari, 1358. Boa Vista,Vitria-ES, Sala 12, CEP:
29075-505. E-mail: carlos.gutierrez@fucape.br.
z
reinaldo@ele.puc-rio.br, DEE-PUC-RJ, Brazil
x
osmani.guillen@bcb.gov.br, Central Bank of Brazil and Ibmec-RJ, Brazil.
1
1 Introduction
In the modeling of economic and nancial time series, the vector autoregressive
(VAR) model has become the standard linear model used in empirical works. An
important aspect of empirical research on the specication of VAR models is deter-
mination of the lag order of the autoregressive lag polynomial, since all inferences
in the VAR model depend on the correct model specication. Several works have
demonstrated the eect of lag length selection. Lutkepohl (1993) indicated that
selecting a higher order lag length than the true one causes an increase in the
mean square forecast errors of the VAR and that undertting the lag length often
generates autocorrelated errors. Braun and Mittnik (1993) showed that impulse
response functions and variance decompositions are inconsistently derived from the
estimated VAR when the lag length diers from the true length. When cointegra-
tion restrictions are considered in the model, the eect of lag length selection on
the cointegration tests has been demonstrated. For example, Johansen (1991) and
Gonzalo (1994) pointed out that VAR order selection can aect proper inference
about cointegrating vectors and rank.
Recently empirical works have considered other kinds of restrictions in the VAR
model (e.g., Engle and Issler, 1995; Caporale, 1997; Mamingi and Sunday, 2003). En-
gle and Kozicki (1993) showed that VAR models can have other types of restrictions,
called common cyclical features, which are restrictions on the short-run dynamics.
These restrictions are dened in the same way as cointegration restrictions, but while
cointegration refers to relations among variables in the long-run, common cyclical
restrictions refer to relations in the short-run. Vahid and Engle (1993) proposed the
serial correlation common feature (SCCF) as a measure of common cyclical features.
SCCF restrictions might be imposed in a covariance stationary VAR model or in a
cointegrated VAR model. The concept of serial correlation common features appears
to be useful. It means that stationary time series move together in a way such that
there are linear combinations of these variables which yield white noise processes
and that their impulse response functions are collinear. In several practical applica-
2
tions the existence of short-run comovements between stationary time series (e.g.,
between rst dierenced cointegrated I(1)) has been analyzed. For instance, Engle
and Issler (1995) found common cycle comovement in U.S. sectoral output data;
Hecq (2002) and Engle and Issler (1993) found common cycles in Latin American
countries; and Carrasco and Gomes (2009) found common international cycles in
GNP data for Mercosur countries.
When short-run restrictions are imposed in cointegrated VAR models, it is pos-
sible to dene a weak version of SCCF restrictions. Hecq, Palm and Urbain (2006)
dened a weak version of SCCF restrictions, which they denominated weak-form
(WF) common cyclical restrictions. A fundamental dierence between SCCF and
WF restrictions is the form in which each one imposes restrictions on the represen-
tation of the vector error correction model (VECM)
1
. When SCCF are imposed, all
matrices of a VECM have rank less than the number of variables analyzed. On the
other hand, with WF restrictions, all matrices except the long-run matrix have rank
less than the number of variables being analyzed. Hence, WF restrictions impose less
constraint on VECM parameters. Some advantages emerge when WF restrictions
are considered. First, due to the fact that the weak form common cyclical method
does not impose constraints on the cointegration space; the rank of common cyclical
features is not limited by the choice of cointegrating rank.
The literature has shown how to select an adequate lag order of a covariance
stationary VAR model and an adequate lag order of a VAR model subject to coin-
tegration restrictions. Among the classical procedures are information criteria, such
as Akaike (AIC), Schwarz (SC) and Hannan-Quinn (HQ) (Lutkepohl, 1993). Kil-
ian (2001) studied the performance of traditional AIC, SC and HQ criteria of a
covariance stationary VAR model. Vahid and Issler (2002) analyzed the standard
information criteria in a covariance stationary VAR model subject to SCCF re-
striction and more recently Guilln, Issler and Athanasopoulos (2005) studied the
standard information criteria in VAR models with cointegration and SCCF restric-
1
When a VAR model has cointegration restriction it can be represented as a VECM. This
representation is also known as Granger Representation Theorem (Engle and Granger 1987).
3
tions. However, when cointegrated VAR models contain an additional weak form
of common cyclical features, there are no reported works on how to appropriately
determine the VAR model order.
The objective of this paper is to investigate the performance of information
criteria in selecting the lag order of a VAR model when the data are generated
from a true VAR with cointegration and WF restrictions, referred to as the correct
model. We carry out the following two procedures: a) the use of standard criteria, as
proposed by Vahid and Engle (1993), referred to here as IC(j), and /) the use of an
alternative model selection criterion (see Vahid and Issler, 2002 and Hecq, Palm and
Urbain, 2006), consisting of simultaneously selecting the lag order p and the number
of weak forms of common cyclical features, :, which is referred to as IC(j. :)
2
. The
most relevant results can be summarized as follows. The information criterion that
selects the pair (j. :) performs better than the model chosen by conventional criteria,
especially the AIC(j. :) criterion. The cost of ignoring additional WF restrictions in
vector autoregressive modeling can be high, particularly when the SC(j) criterion
is used.
The rest of this work is organized as follows. Section 2 presents the economet-
ric model. Section 3 discusses the information criteria. Section 4 shows a Monte
Carlo simulation and Section 5 presents the results. Finally, Section 6 contains our
conclusions.
2 The Econometric Model
We show the VAR model with short-run and long-run restrictions. First, we consider
a Gaussian vector autoregression of nite order j, called VAR(j), such that:

t
=
p

i=1

ti

t
(1)
where,
t
is a vector of : rst order integrated series, 1(1),
i
, i = 1. . . . . j are
matrices of dimension : :,
t
~ `o::c| (0. !) . 1 (
t
) = 0 and 1 (
t

) ={! , if
2
This is quite recent in the literature (see, Hecq et al., 2006).
4
t = t and 0
nn
, if t ,= t, where ! is nonsingular}. The model (1) can be written
equivalently as; H(1)
t
=
t
where 1 represents the lag operator and H(1) =
1
n


p
i=1

i
1
i
such that when 1 = 1, H(1) = 1
n


p
i=1

i
. If cointegration is
considered in (1) the (: :) matrix H() satises two conditions: a) rank (H(1)) =
:, 0 < : < :, such that H(1) can be expressed as H(1) = c,
0
, where c and ,
are (: :) matrices with full column rank, :; and b) the characteristic equation
[H(1)[ = 0 has :: roots equal to 1 and all other are outside the unit circle. These
assumptions imply that
t
is cointegrated of order (1. 1). The elements of c are the
adjustment coecients and the columns of , span the space of cointegration vectors.
We can represent a VAR model as a Vector Error Correction Model (VECM). By
decomposing the polynomial matrix H(1) = H(1) 1H

(1) ^, where ^ = (1 1)
is the dierence operator, a VECM is obtained:
^
t
= c,
0

t1

p1

i=1
I
i
^
ti

t
(2)
where: c,
0
= H(1), I
j
=

p
k=j+1

k
for , = 1. ..... j 1 and I
0
= 1
n
. The
VAR(j) model can include additional short-horizon restrictions as shown by Vahid
and Engle (1993). We consider an interesting WF restriction (as dened by Hecq,
Palm and Urbain, 2006) that does not impose constraints on long-run relations.
Denition 1 The weak form (WF) holds in (2) if, in addition to cointegration
restriction, there exists an (: :) matrix

, of rank :, whose columns span the
cofeature space, such that

,
0
(^
t
c,
t1
) =

,
0

t
. where

,
0

t
is an s-dimensional
vector that constitutes an innovation process with respect to information prior to
period t, given by
t1
.
t2
. ....
1
.
Equivalent to denition 1, we consider WF restrictions in the VECM if there
exists a cofeature matrix

, that satises the following assumption:
Assumption 1 :

,
0
I
j
= 0
sn
for , = 1. ..... j 1.
Therefore this is a naturally weaker alternative assumption which implies that
the common cyclical part is reduced to white noise by taking a linear combination of
5
the variables in the rst dierences adjusted for long-run eects. Imposing WF re-
strictions is convenient because it allows studying of both cointegration and common
cyclical feature without the constraint
3
: : _ :.
We can rewrite the VECM with WF restrictions as a model of reduced-rank
structure. In (2) let A
t1
= [^
0
t1
. .....^
0
tp+1
[
0
and 4 = [I
1
. ..... I
p1
[. Therefore,
we obtain:
^
t
= c,
t1
4A
t1

t
(3)
If assumption (1) holds, then matrices I
i
. i = 1. .... j are all of rank (: :) and we
can write 4 =

,
?
d =

,
?
[d
1
. ..... d
p1
[, where,

,
?
is : (: :) full column rank
matrix, d has dimension (: :) :(j 1). and the matrices d
i
. i = 1. .... j 1 all
have rank (: :) :. Hence, given assumption (1), there exists

, of : : such
that

,
0

,
?
= 0. That is,

,
?
: (: :) is a full column rank orthogonal to the
complement of

, with :c:/(

,.

,
?
) = :. Rewriting model (3) we have:
^
t
= c,
t1


,
?
(d
1
. d
2
. .... d
p1
) A
t1

t
(4)
= c,
t1


,
?
dA
t1

t
(5)
Estimation of (5) is carried out via the switching algorithms (see Centoni et al.,
2007; Hecq, 2006) that use the procedure in estimating reduced-rank regression
models as suggested by Anderson (1951). There is a formal connection between a
reduced-rank regression and the canonical analysis, as noted by Izenman (1975),
Box and Tiao (1977), Tso (1981) and Velu Reinsel and Wichern (1986). When all
the matrix coecients of the multivariate regression are full rank, it can be esti-
mated by the usual least squares or maximum likelihood procedures. But when
the matrix coecients are of reduced-rank they have to be estimated using the
reduced-rank regression models of Anderson (1951). The use of canonical analysis
may be regarded as a special case of reduced-rank regression. More specically, the
maximum-likelihood estimation of the parameters of the reduced-rank regression
3
Since the SCCF also imposes constraints on the long-run matrix
0
= (1), which has
dimension n, the cointegration restrictions, r, and SCCF restrictions, s, must satisfy r + s _ n.
6
model may result in solving a problem of canonical analysis
4
. Therefore, we can use
the expression Cc:Co::A
t
. 2
t
[A
t1
which denotes the partial canonical correla-
tions between A
t
and 2
t
: both sets concentrate out the eect of A
t1
allowing us
to obtain canonical correlation (see Johansen, 1995), represented by the eigenvalues

`
1


`
2


`
3
.......

`
n
. The Johansen test statistic is based on canonical cor-
relation. In model (2) we can use the expression Cc:Co::^
t
.
t1
[A
t1
where
A
t1
= [^
0
t1
. .....^
0
tp+1
[
0
which summarizes the reduced-rank regression proce-
dure used in the Johansen approach. This means that we extracts the canonical
correlations between ^
t
and
t1
: both sets concentrate out the eect of lags of
A
t1
. In order to test for the signicance of the : largest eigenvalues, we can rely
on Johansens trace statistic (6):

r
= 1
n

i=r+1
1:(1

`
2
i
) i = 1. .... : (6)
where the eigenvalues 0 <

`
n
< ... <

`
1
are the solution of : [`:
11
:
1
10
:
00
:
01
[ =
0, where :
ij
. i. , = 0.1. are the second moment matrices: :
00
=
1
T

T
t=1
n
0t
n
0
0t
,
:
10
=
1
T

T
t=1
n
1t
n
0
0t
, :
01
=
1
T

T
t=1
n
0t
n
0
1t
, :
11
=
1
T

T
t=1
n
1t
n
0
1t
of the residuals n
0t
and n
1t
obtained in the multivariate least squares regressions ^
t
= (^
t1
,...,^
tp+1
)
n
0t
and
t1
= (^
t1
. ...^
tp+1
)n
1t
respectively (see, Hecq et al., 2006; Johansen,
1995). The result of the Johansen test is a superconsistent estimation of ,. More-
over, we could also use a canonical correlation approach to determine the rank of
the common features space due to WF restrictions. This is a test for the existence of
cofeatures in the form of linear combinations of the variables in the rst dierences,
corrected for long-run eects which are white noise (i.e.,

,
0
(^
t
c,
t1
) =

,
0

t
where

,
0

t
is a white noise). We use canonical analysis is this work to estimate, and
select the lag-rank of VAR models, as shown in next sections.
4
This estimation is referred to as full information maximum likelihood - FIML. (see Johansen,
1995).
7
3 Model Selection Criteria
In model selection we use two procedures to identify the VAR model order: the
standard selection criteria, IC(j), and the modied informational criteria, IC(j. :),
a novelty in the literature, which consists of identifying j and : simultaneously.
The model estimation following the standard selection criteria, IC(j), originally
used by Vahid and Engle (1993) entails the following steps:
1. Estimate j using standard informational criteria: Akaike (AIC),
Schwarz (SC) and Hanna-Quinn (HQ). We chose the lag length of the
VAR in levels that minimizes the information criteria.
2. Using the lag length chosen in the previous step, nd the number of
cointegration vectors, :. using the Johansen cointegration test
5
.
3. Conditional on the results of the cointegration analysis, estimate a
nal VECM then calculate the multi-step ahead forecast.
The above procedure is followed when there is evidence of cointegration restric-
tions. We check the performance of IC(j) when WF restrictions are imposed on the
true model. Additionally we check the performance of IC(j. :) alternative selection
criteria. Vahid and Issler (2002) analyzed a covariance-stationary VAR model with
SCCF restrictions. They showed that the use of IC(j. :) performs better than IC(j)
in VAR model lag order selection. In the present work we analyze the cointegrated
VAR model with WF restrictions in order to analyze the performance of IC(j) and
IC(j. :) for model selection. The question investigated is: Does IC(j. :) perform
better than IC(j)? This is an important question we aim to answer in this work.
The procedure to choose the lag order and the rank of the structure of short-run
restrictions is carried out by minimizing the following modied information criteria
(see; Vahid and Issler, 2002; Hecq, 2006).
1C (j. :) =
T

i=ns+1
ln(1 `
2
i
(j))
2
1
N (7)
5
Cointegration rank and vectors are estimated using the FIML, as shown in Johansen (1991).
8
HQ(j. :) =
T

i=ns+1
ln(1 `
2
i
(j))
2 ln(ln 1)
1
N (8)
oC(j. :) =
T

i=ns+1
ln(1 `
2
i
(j))
ln 1
1
N (9)
N = [: (: (j 1)) : :[ [: (: (j 1) (: :))[
where : is the number of variables in model (2) and ` is a number of parameters.
` is obtained by subtracting the total number of mean parameters in the VECM
(i.e., :
2
(j1)::), for given : and j, from the number of restrictions the common
dynamics imposes from :(:(j1)):(::). The eigenvalues `
i
are calculated
for each j. In order to calculate the pair (j. :) we assume that no restriction exists,
that is, : = : (see Hecq, 2006). We x j in model (3) and then nd `
i
i = 1. 2...: by
computing the cc:co::(^
t
. A
t1
[
t1
). This procedure is followed for every j and
in the end we choose the j and : that minimize the IC(j. :). After selecting the pair
(j. :) we can test the cointegration relation using the procedure of Johansen. Finally
we estimate the model using the switching algorithms as shown in the next section.
Notice that in this simultaneous selection, testing the cointegration relation is the
last procedure followed, so we are inverting the hierarchical procedure followed by
Vahid and Engle (1993) where the rst step is to select the number of cointegration
relations. This may be an advantage, especially when : is over-estimated. Few works
have analyzed the order of VAR models considering modied IC(j. :). As mentioned,
Vahid and Issler (2002) suggested the use of IC(j. :) to simultaneously choose the
order j and a number of reduced rank structure : in a covariance stationary VAR
model subject to SCCF restrictions. However, no work has analyzed the order of
the VAR model with cointegration and WF restrictions using a modied criterion,
which is exactly the contribution of this paper.
To estimate the VAR model considering cointegration and WF restrictions we
use the switching algorithms model as considered by Hecq (2006). Consider the
9
VECM given by:
^
t
= c,
0

t1


,
?
dA
t1

t
(10)
A full description of switching algorithms is presented below in four steps:
otcj1 : Estimation of the cointegration vectors ,.
Using the optimal pair ( j. :) chosen by the information criteria (7), (8)
or (9), we estimate , (and so its rank, : = :) using the Johansen coin-
tegration test.
otcj2 : Estimation of

,
?
and d.
Taking the estimate of , in step one, we proceed to estimate

,
?
and d.
Hence, we run a regression of ^
t
and of A
t1
on

,
0

t1
. We label the
residuals n
0
and n
1
, respectively. Therefore, we obtain a reduced rank
regression:
n
0t
=

,
?
dn
1t

t
(11)
where d can be written as d =
_
C
1
. .... C
( p1)
_
of (: :) :( j 1) and

,
?
of : (: :). We estimate (11) by FIML. Thus, we can obtain

,
?
and

d.
otcj3 : Estimation of the maximum likelihood (ML) function.
Given the parameters estimated in steps 1 and 2 we use a recursive
algorithm to estimate the maximum likelihood (ML) function. We cal-
culate the eigenvalues associated with

d,

`
2
i
i = 1. .... : and the matrix
of residuals

max
r; s= s
. Hence, we compute the ML function:
1
0
max; r<n; s= s
=
1
2
_
ln

max

r<n; s= s

i=1
ln
_
1

`
2
i
_
_
(12)
If : = :, instead of (12) we use the derived log-likelihood: 1
max; r=n; s= s
=

T
2
ln

max
r=n; s= s

. The determinant of the covariance matrix for : = :

cointegration vector is calculated by
ln

max

r=n; s= s

= ln

:
00
:
01
:
1
11
:
10

i=1
ln
_
1

`
2
i
_
(13)
10
where :
ij
refers to cross moment matrices obtained in multivariate least
square regressions from ^
t
and A
t1
on
t1
. In this case, estimation
does not entail using an iterative algorithm yet because the cointegrating
space spans 1
n
.
otcj4 : Reestimation of ,.
We reestimate , to obtain a more appropriate value for the parameters.
In order to reestimate , we compute the Cc:Co::
_
^
t
.
t1
[

dA
t1
_
and thus using the new

, we can repeat step 2 to reestimate

,
?
and
d. Then, we calculate the new value of the ML function in the step
3. Hence, we obtain 1
1
max; r= r; s= s
to calculate ^1 = (1
1
max; r= r; s= s
-
1
0
max; r= r; s= s
).
We repeat steps 1 to 4 to choose

,
?
and d until convergence is reached ( i.e.,
^1 < 10
7
). In the end, the optimal parameters j, : and : are obtained and they
can be used to estimate and forecast of a VECM with WF restrictions.
4 Monte-Carlo Design
The simple real business cycle models and also the simplest closed economy monetary
dynamic stochastic general equilibriummodels are three-dimensional. Consumption,
saving and output and prices, output and money are notable examples. Motivated
by these applications and according the previous work of Vahid and Issler (2002) we
construct a Monte-Carlo experiment in a three-dimensional environment. Therefore,
the data generating processes considering a VAR model with three variables, one
cointegration vector, and two cofeatures vectors (i.e., : = 3, : = 1 and : = 2,
11
respectively). , and

, satisfy:
, =
_
_
1.0
0.2
1.0
_
_
.

, =
_
_
1.0 0.1
0.0 1.0
0. 0.
_
_
_
_

1t

2t

3t
_
_
s `
_
_
_
_
0
0
0
_
_
.
_
_
1.0 0.0 0.0
0.0 1.0 0.0
0.0 0.0 1.0
_
_
_
_
Consider the VAR(3) model:
t
=
1

t1

2

t2

3

t3

t
. The VECM
representation as a function of the VAR level parameters can be written as:
^
t
= (
1

2

3
1
3
)
t1
(
2

3
)^
t1

3
^
t2

t
(14)
The VAR coecients must simultaneously obey the restrictions: a) The cointe-
gration restrictions: c,
0
= (
1

2

3
1
3
) ; b) WF restrictions:

,
0

3
= 0
(iii)

,
0
(
2

3
) = 0 and c) the covariance-stationary condition. Considering the
cointegration restrictions, we can rewrite (14) as the following VAR(1):

t
= 1
t1

t
(15)

t
=
_
_

t

t1
,
0

t
_
_
. 1 =
_
_
(
2

3
)
3
c
1
3
0 0
,(
2

3
) ,
0

3
,
0
c 1
_
_
and
t
=
_
_

t
0
,
0

t
_
_
Thus, equation (15) will be covariance-stationary if all eigenvalues of matrix
1 lie inside the unit circle. An initial idea to design the Monte-Carlo experiment
can consist of constructing the companion matrix (1) and verifying whether the
eigenvalues of the companion matrix all lie inside the unit circle. This can be
carried out by selecting their values from a uniform distribution, and then verifying
whether or not the eigenvalues of the companion matrix all lie inside the unit circle.
However, this strategy could lead to a wide spectrum to search for adequate values
for the companion matrix. Hence, we follow an alternative procedure. We propose
an analytical solution to generate a covariance-stationary VAR, based on the choice
of the eigenvalues, and then on the generation of the respective companion matrix.
In the appendix we present a detailed discussion of the nal choice of these free
parameters, including analytical solutions. In our simulation, we constructed 100
12
data generating processes and for each of these we generate 1000 samples containing
1000 observations. To reduce the impact of the initial values, we considered only the
last 100 and 200 observations. All the experiments were conducted in the MatLab
environment.
5 Results
Figure 1 shows one example of the three-dimensional VAR model with cointegration
and WF restrictions for 100 and 200 observations.
0 1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 0 9 0 1 0 0
- 1 0 0
- 9 0
- 8 0
- 7 0
- 6 0
- 5 0
- 4 0
- 3 0
- 2 0
T im e
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0 1 6 0 1 8 0 2 0 0
- 1 0 0
- 8 0
- 6 0
- 4 0
- 2 0
0
T im e
Figure 1. One example of a VAR(3) model with n=3,r=1 and s=2 for 100 and 200 observations
The values in Table 1 represent the percentage of time that the model selection
criterion, IC(j), chooses the cell corresponding to the lag and number of cointegra-
tion vectors in 100,000 runs. The true lag-cointegrating vectors are identied by
boldface numbers and the selected lag-cointegration vectors chosen the most times
by the criterion are underlined. In Table 1, the results show that, in general, the
AIC most often chooses the correct lag length for 100 and 200 observations. For
13
example, for 100 observations, the AIC, HQ and SC criteria chose the true lag, j,
54.08%, 35.62% and 17.48% of the times respectively. Note that all three criteria
chose the correct rank of cointegration (: = 1). When 200 observations was consid-
ered, the correct lag length was chosen 74.72%, 57.75% and 35.28% of the times for
AIC, HQ and SC respectively. Again, all three criteria selected the true cointegrated
rank : = 1. Table 2 contains the percentage that the alternative model selection
criterion, IC(j. :), chooses that cell, corresponding to the lag-rank and number of
cointegrating vectors in 100,000 runs. The true lag-rank-cointegration vectors are
identied by boldface numbers and the best lag-rank combination chosen the most
times by each criterion are underlined. In Table 2, the results show that, in general,
the AIC(j. :) criterion more frequently chooses the lag-rank for 100 and 200 obser-
vations. For instance, for 100 observations, the AIC(j. :), HQ(j. :) and SC(j. :)
criteria more often choose the true pair (j. :) = (3. 1), 56.34%, 40.85% and 25.2% of
the times, respectively. For 200 observations, the AIC(j. :), HQ(j. :) and SC(j. :)
criteria more frequently choose the true pair (j. :) = (3. 1), 77.06%, 62.58% and
45.03% of the times respectively. Note that all three criteria more often choose
the correct rank of cointegration (: = 1) in both samples. What happened when
the weak form common cyclical restrictions are ignored? Tables 1 and 2 also show
the relative performance of IC(j. :) vis--vis IC(j). For instance, for 1 = 100 the
SC(j. :) has a success rate of 25.2% in selecting the true j = 3, while the SC(j)
only has a success rate of 17.48%. This represents a gain of more than 44%. For
1 = 200, the gains are more than 27%. For 1 = 100 the HQ(j. :) selects the true
j =3 with 40.85% accuracy while the HQ(j) only has a success rate of 35.62%. This
represents a gain of 14%. For 1 = 200, the gains are more than 8%. For 1 = 100
the AIC(j. :) has a success rate of 56.34% in choosing the true j = 3, in comparison
with a rate of 54.08% for the AIC(j), a gain of more than 4%. For T=200, the gains
are more than 3%. Thus, it appears that when using the AIC(p,s) criteria the cost
of ignoring the weak form common cyclical restriction is low.
The most relevant results can be summarized as follows:
14
All criteria (AIC, HQ and SC) choose the correct parameters more
often when using IC(j. :) vis--vis IC(j).
There is a cost of ignoring additional weak form common cyclical re-
strictions in the model especially when the SC(j) criterion is used. In
general, the standard Schwarz, SC(j), or Hannan-Quinn, HQ(j), selec-
tion criteria should not be used for this purpose in small samples due to
the tendency to identify an under-parameterized model.
The AIC performs better in selecting the true model more frequently
for both the IC(j. :) and the IC(j) criteria.
6 Conclusions
In this work, we considered an additional weak form restriction of common cycli-
cal features in a cointegrated VAR model in order to analyze the appropriate way
to choose the correct lag order. These additional WF restrictions are dened in
the same way as cointegration restrictions. While cointegration refers to relations
among variables in the long-run, the common cyclical restrictions refer to relations
in the short-run. Two methodologies have been used for selecting lag length; the
traditional information criterion, IC(j), and an alternative criterion (IC(j. :)) that
selects the lag order j and the rank structure : due to the weak form common
cyclical restrictions.
The results indicate that the information criterion that selects the lag length
and the rank order performs better than the model chosen by conventional criteria.
When the WF restrictions are ignored there is a non trivial cost in selecting the
true model with standard information criteria. In general, the standard Schwarz
or Hannan-Quinn selection criteria should not be used for this purpose in small
samples, due to the tendency to identify an under-parameterized model.
In applied work, when the VAR model contains WF and cointegration restric-
tions, we suggest the use of AIC(j. :) criteria to choose the lag-rank, since it provides
15
considerable gains in selecting the correct VAR model. Since no work in the litera-
ture has analyzed a VAR model with WF common cyclical restrictions, the results of
this work provide new insights and incentives to proceed with this kind of empirical
work.
References
Anderson, T. W, (1951) Estimating linear restrictions on regression coecients
for multivariate normal distributions. Ann. Math. Statist. 22 327-351. [Correction
(1980) Ann. Statist. 8 1400].
Box GE, Tiao GC. (1977). A canonical analysis of multiple time series. Bio-
metrika, 64:355-365.
Braun PA, Mittnik S. (1993). Misspecications in Vector Autoregressions and
Their Eects on Impulse Responses and Variance Decompositions. Journal of Econo-
metrics, 59, 319-41.
Caporale GM. (1997). Common Features and Output Fluctuations in the United
Kingdom. Economic-Modelling, 14: 1-9.
Centoni M, Cubadda G, Hecq. (2007) Common Shocks, Commo Dynamics, and
the International Business Cycle. Economic-Modelling, 24, 149-166.
Carrasco CG, Gomes F. (2009). Evidence on Common Features and Business
Cycle Synchronization in Mercosur. Brazilian Review of Econometrics. Forthcom-
ing.
Engle R, Granger C. (1987). Cointegration and error correction: representation,
estimation and testing. Econometrica, , 55, 251-76.
Engle R, Kozicki S.(1993). Testing for common features. Journal of Business
and Economic Statistics, 11: 369-395.
16
Engle R, Issler JV. (1995). Estimating Common Sectoral Cycles. Journal of
Monetary Economics, 35:83-113.
Engle R, Issler JV.(1993). Common Trends and Common Cycles in Latin Amer-
ica. Revista Brasileira de Economia, vol. 47, n
o
2, pp. 149-176, 1993.
Gonzalo, J. (1994). Five Alternative Methods of Estimating Long-Run Equilib-
rium Relationships. Journal of Econometrics, 60 issue 1-2, 203-233.
Guilln, Issler, Athanasopoulos. (2005). Forecasting Accuracy and Estimation
Uncertainty using VAR Models with Short- and Long-Term Economic Restrictions:
A Monte-Carlo Study. Ensaios Econmicos EPGE, 589.
Hecq A. (2006). Cointegration and Common Cyclical Features in VAR Models:
Comparing Small Sample Performances of the 2-Step and Iterative Approaches.
Mimeo.
Hecq A, Palm FC, Urbain JP. (2006). Testing for Common Cyclical Features in
VAR Models with Cointegration, Journal of Econometrics, 132: 117-141.
Hecq, A. (2002), Common Cycles and Common Trends in Latin America. Medium
Econometrische Toepassingen, 10, 20-25.
Izenman AJ. (1975). Reduced rank regression for the multivariate linear model,
Journal of. multivariate Analysis, 5: 248-264.
Johansen, S. (1991). Estimation and Hypothesis Testing of Cointegration Vectors
in Gaussian Vector Autoregressive Models. Econometrica, 59, 1551-1580.
Johansen S. (1995). Likelihood-based Inference in Cointegrated Vector Autore-
gressive Models. Oxford Univ. Press. Z.
Kilian L. (2001). Impulse response analysis in vector autoregressions with un-
known lag order. Journal of Forecasting, 20: 161-179.
17
Ltkepohl H. (1993). Introduction to Multiple Time Series Analysis. Springer,
New York.
Mamingi N, Iyare SO. (2003). Convergence and Common Features in Interna-
tional Output: A Case Study of the Economic Community of West African States,
1975-1997. Asian-African-Journal-of-Economics-and-Econometrics, 3: 1-16.
Tso MK. (1981). Reduced rank regression and Canonical analysis. Journal of
the Royal Statistical Society, 43: 183-189.
Vahid F, Engle R. (1993). Common trends and common cycles. Journal of
Applied Econometrics, 8: 341-360.
Vahid F, Issler JV. (2002). The Importance of Common Cyclical Features in
VAR Analysis: A Monte Carlo Study. Journal of Econometrics, 109: 341-363.
Velu RP, Reinsel GC, Wichern DW. (1986). Reduced rank models for multiple
times series. Biometrika, 73: 105-118.
18
Appendix A: Tables
0 1 2 3 0 1 2 3
Selected Lag
1 0.000 0.996 0.359 0.031 0.000 0.095 0.016 0.003
2 0.002 32.146 1.136 0.048 0.000 17.073 0.686 0.033
3 2.792 54.082 0.902 0.041 0.012 74.721 1.488 0.108
AIC(p) 4 0.737 4.068 0.091 0.003 0.005 4.177 0.081 0.006
5 0.392 0.987 0.031 0.000 0.013 0.828 0.020 0.000
6 0.219 0.333 0.014 0.000 0.023 0.257 0.005 0.000
7 0.166 0.173 0.006 0.000 0.039 0.133 0.002 0.000
8 0.133 0.107 0.005 0.000 0.060 0.115 0.001 0.000
1 0.000 3.884 1.915 0.165 0.000 1.098 0.243 0.021
2 0.002 52.593 1.907 0.080 0.000 37.390 1.614 0.098
3 2.600 35.617 0.612 0.027 0.012 57.749 1.146 0.082
HQ(p) 4 0.065 0.189 0.007 0.000 0.001 0.158 0.004 0.000
5 0.059 0.037 0.000 0.000 0.009 0.082 0.001 0.000
6 0.073 0.025 0.000 0.000 0.016 0.076 0.000 0.000
7 0.059 0.019 0.001 0.000 0.030 0.070 0.000 0.000
8 0.053 0.011 0.000 0.000 0.044 0.055 0.001 0.000
1 0.000 8.344 6.609 0.511 0.000 3.964 1.385 0.093
2 0.003 61.966 2.279 0.105 0.000 55.156 2.776 0.169
3 2.042 17.485 0.313 0.015 0.012 35.283 0.728 0.044
SC(p) 4 0.049 0.045 0.000 0.000 0.001 0.083 0.002 0.000
5 0.071 0.025 0.000 0.000 0.007 0.076 0.001 0.000
6 0.057 0.016 0.000 0.000 0.013 0.063 0.000 0.000
7 0.036 0.009 0.000 0.000 0.025 0.056 0.000 0.000
8 0.017 0.003 0.000 0.000 0.027 0.035 0.001 0.000
Numbers represent the percentage times that the model selection criterion choice that cell corresponding to the lag and
number of cointegration vectors in 100,000 realizations. The true lag-cointegrating vectors are indentified by bold numbers.
Table I Performance of information criterion, IC(p ), in selecting the lag order p
Number of observations = 100
Selected cointegrated vectors
Number of observations = 200
Selected cointegrated vectors
Frequency of lag(p) and cointegrating vectors (r) choice by different criteria for trivariate VAR model in
levels when the true model have parameters: p = 3 and r = 1.
19
Johansen Tested coint. Vectors (r) 0 1 2 3
Selected rank (s) 1 2 3 1 2 3 1 2 3 1 2 3
Selected lag (p)
Sample size = 100
1 - - - - - - - - - - - -
2 0.002 0.000 0.000 39.049 0.001 0.000 1.218 0.000 0.000 0.056 0.000 0.000
3 0.301 0.000 0.000 56.341 0.003 0.000 1.559 0.000 0.000 0.053 0.000 0.000
AIC(p ,s ) 4 0.004 0.000 0.000 1.186 0.001 0.000 0.070 0.000 0.000 0.001 0.000 0.000
5 0.000 0.000 0.000 0.114 0.001 0.000 0.012 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.020 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.006 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1 - - - - - - - - - - - -
2 0.002 0.000 0.000 55.563 0.000 0.000 1.888 0.000 0.000 0.081 0.000 0.000
3 0.267 0.000 0.000 40.855 0.000 0.000 1.207 0.000 0.000 0.043 0.000 0.000
HQ(p ,s ) 4 0.000 0.000 0.000 0.088 0.000 0.000 0.005 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1 - - - - - - - - - - - -
2 0.004 0.000 0.000 70.971 0.000 0.000 2.574 0.000 0.000 0.113 0.000 0.000
3 0.221 0.000 0.000 25.204 0.000 0.000 0.887 0.000 0.000 0.025 0.000 0.000
SC(p ,s ) 4 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Sample size = 200
1 - - - - - - - - - - - -
2 0.000 0.000 0.000 18.797 0.000 0.000 0.681 0.000 0.000 0.038 0.000 0.000
3 0.000 0.000 0.000 77.065 0.002 0.000 2.260 0.000 0.000 0.145 0.000 0.000
AIC(p ,s ) 4 0.000 0.000 0.000 0.908 0.000 0.000 0.035 0.000 0.000 0.001 0.000 0.000
5 0.000 0.000 0.000 0.063 0.000 0.000 0.002 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.003 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1 - - - - - - - - - - - -
2 0.000 0.000 0.000 33.952 0.000 0.000 1.370 0.000 0.000 0.086 0.000 0.000
3 0.000 0.000 0.000 62.576 0.000 0.000 1.877 0.000 0.000 0.111 0.000 0.000
HQ(p ,s ) 4 0.000 0.000 0.000 0.027 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1 - - - - - - - - - - - -
2 0.000 0.000 0.000 50.983 0.000 0.000 2.351 0.000 0.000 0.146 0.000 0.000
3 0.000 0.000 0.000 45.028 0.000 0.000 1.416 0.000 0.000 0.076 0.000 0.000
SC(p ,s ) 4 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Table II. Performance of information criterion, IC(p,s), in selecting p and s simultaneously.
Numbers represent the percentage times that the simultaneous model selection criterion IC(p,s) choice that cell, corresponding to the lag-rank and number of cointegrating vectors in
100,000 realizations. The true lag-rank-cointegration vectors are identified by bold numbers and the best lag-rank-cointegration vectors chosen by criterias are identified by underline
lines.
20
Appendix B: VAR Restrictions for the DGPs
Consider the vector autoregressive, VAR(3), model :

t
=
1

t1

2

t2

3

t3

t
(16)
with parameters:
1
=
_
_
c
1
11
c
1
12
c
1
12
c
1
21
c
1
22
c
1
22
c
1
31
c
1
32
c
1
32
_
_
,
2
=
_
_
c
2
11
c
2
12
c
2
12
c
2
21
c
2
22
c
2
22
c
2
31
c
2
32
c
2
32
_
_
and
3
=
_
_
c
3
11
c
3
12
c
3
12
c
3
21
c
3
22
c
3
22
c
3
31
c
3
32
c
3
32
_
_
. We consider the cointegration vectors , =
_
_
,
11
,
21
,
31
_
_
, the cofea-
tures vectors

, =
_
_

,
11

,
12

,
21

,
22

,
31

,
32
_
_
and the adjustament matrix c =
_
_
c
11
c
21
c
31
_
_
. The
long-run relation is dened by c,
0
= (
1

2

3
1
3
). Thus, the VECM respre-
sentation is:
^
t
= c,
0

t1
(
2

3
)^
t1

3
^
t2

t
(17)
We can rewrite equation (17) as a VAR(1):

t
= 1
t1

t
(18)
where
t
=
_
_

t

t1
,
0

t
_
_
. 1 =
_
_
(
2

3
)
3
c
1
3
0 0
,(
2

3
) ,
0

3
,
0
c 1
_
_
and
t
=
_
_

t
0
,
0

t
_
_
1) Short-run restrictions (WF)
We now impose the common cyclical restrictions (i) and (ii) on model (16).
Let, G = [1
21
1 1
31
[, 1 = [(1
32
1
31
),(1
21
1
22
)[, 1
j1
=

,
j1
,

,
11
, 1
j2
=

,
j2
,

,
12
(, = 2. 3) and o = ,
11
G ,
21
1 ,
31
(i)

,
0

3
= 0 ==
3
=
_
_
Gc
3
31
Gc
3
32
Gc
3
33
1c
3
31
1c
3
32
1c
3
33
c
3
31
c
3
32
c
3
33
_
_
(ii)

,
0
(
2

3
) = 0 ==

,
0

2
= 0 ==
2
=
_
_
Gc
2
31
Gc
2
32
Gc
2
33
1c
2
31
1c
2
32
1c
2
33
c
2
31
c
2
32
c
2
33
_
_
2) Long-run restrictions (cointegration)
21
The cointegration restrictions are specied by (iv) and (v):
(i) ,
0
(
2

3
) = [(c
2
31
c
3
31
)o (c
2
32
c
3
32
)o (c
2
33
c
3
33
)o[ and
,
0

3
= [c
3
31
o c
3
32
o c
3
33
o[
() ,
0
c1 = , =
_
,
11
,
21
,
31

_
_
c
11
c
21
c
31
_
_
1 = ,
11
c
11
,
21
c
21
,
31
c
31
1
Taking into account the short- and long-run restrictions, the companion matrix
1 can be represented as:
1 =
_
_
(
2

3
)
3
c
1
3
0 0
,(
2

3
) ,
0

3
,
0
c 1
_
_
=
_

_
G(c
2
31
c
3
31
) G(c
2
32
c
3
32
) G(c
2
33
c
3
33
) Gc
3
31
Gc
3
32
Gc
3
33
c
11
1(c
2
31
c
3
31
) G(c
2
32
c
3
32
) G(c
2
33
c
3
33
) 1c
3
31
1c
3
32
1c
3
33
c
21
(c
2
31
c
3
31
) G(c
2
32
c
3
32
) (c
2
33
c
3
33
) c
3
31
c
3
32
c
3
33
c
31
1 0 0 0 0 0 0
0 1 0 0 0 0 0
0 0 1 0 0 0 0
(c
2
31
c
3
31
)o (c
2
32
c
3
32
)o (c
2
33
c
3
33
)o c
3
31
o c
3
32
o c
3
33
o /
_

_
with / = ,
0
c 1 = ,
11
c
11
,
21
c
21
,
31
c
31
1
3) Covariance-stationary restrictions
Equation (18) will be covariance-stationary if all eigenvalues of matrix 1 lie
inside the unit circle. That is, eigenvalue of matrix 1 is a number ` such that:
[1 `1
7
[ = 0 (19)
The solution of (19) is:
`
7
!`
6
O`
5
d`
4
= 0 (20)
where the parameters !, O, and d are: ! = G(c
2
31
c
3
31
)1(c
2
32
c
3
32
)c
2
33
c
3
33
/,
O = Gc
3
31
1c
3
32
(c
2
33
c
3
33
)/ G/(c
2
31
c
3
31
) 1/(c
2
32
c
3
32
) c
31
o(c
2
33
c
3
33
)
oc
21
(c
2
32
c
3
32
) oc
11
(c
2
31
c
3
31
) c
3
33
and d = c
3
33
/ Gc
3
31
/ 1c
3
32
/ c
31
c
3
33
o
c
3
32
oc
21
c
3
31
oc
11
, and the rst four roots are `
1
= `
2
= `
3
= `
4
= 0.We calculated
22
the parameters of matrices
1
,
2
and
3
as functions of roots (`
5
. `
6
and `
7
) and
free parameters. Hence we have three roots satisfying equation (20)
`
3
!`
2
O` d = 0 (21)
for `
5
, we have: `
3
5
!`
2
5
O`
5
d = 0 ..................................11
for `
6
, we have: `
3
6
!`
2
6
O`
6
d = 0 ..................................12
for `
7
, we have: `
3
7
!`
2
7
O`
7
d = 0 ..................................13
Solving equations 1, 2 and 3 yields: ! = `
7
`
6
`
5
, O = `
6
`
7
`
6
`
5
`
5
`
7
and d = `
5
`
6
`
7
. Equaling these parameters with the relations above we have:
c
2
31
= (1c
2
32
1c
2
32
/ c
31
oc
2
33
`
6
`
7
`
6
`
7
c
2
33
/ `
5
`
6
`
7
/ `
5
`
7

`
5
`
6
c
2
33
oc
2
32
c
21
`
5
),(oc
11
G G/)
c
3
32
= (o
2
`
7
c
11
c
31
/
2
`
7
G`
6
G/
2
/`
7
oc
11
`
6
oc
11
/c
3
31
oc
11
Gc
3
31
o
2
c
2
11

Gc
3
31
/oc
11
`
5
G/
2
`
5
oc
11
/`
7
`
6
c
31
oG`
7
`
5
c
31
oGo
2
c
11
`
5
c
31
o
2
c
11
`
6
c
31

o`
5
G/c
31
oc
31
`
6
G/`
5
`
7
`
6
G`
6
`
7
G/`
5
`
7
G/`
5
`
6
G/oG/
2
c
31
o
2
c
11
/c
31

o
2
c
11
c
31
c
2
33
o
2
c
2
31
c
2
33
GoG
2
c
3
31
c
31
oc
11
c
2
33
/ G/
3
oc
11
/
2
o
2
c
11
1c
2
32
c
31

o
2
c
11
c
31
Gc
3
31
o
2
c
2
32
c
21
Gc
31
oc
2
32
c
21
G/oc
31
G
2
c
3
31
/oc
31
c
2
33
G/oc
11
1c
2
32
/
o`
7
G/c
31
`
5
`
6
c
31
oG`
5
`
7
`
6
c
31
oG`
5
`
7
`
6
oc
11
),(oc
11
1c
31
1Gc
31
/Gc
21

1c
31
G/ oc
11
c
21
Gc
21
),o
c
3
33
= (1/
3
G`
5
G/
2
1oc
11
`
6
1`
7
`
5
1/`
7
oc
11
1/
2
`
7
Go
2
c
21
`
7
c
11

`
6
G/oc
21
oc
21
`
7
G/`
6
G/
2
1`
6
oc
11
1/`
6
o
2
c
11
c
21
`
5
G/oc
21
`
5
oc
11
1/
`
5
o
2
c
11
c
21
`
7
`
6
oc
21
G1/`
7
`
6
G1/`
7
`
5
G1/`
5
`
6
G`
7
`
6
1G`
5
o
2
c
11
c
21
1c
2
32

o
2
c
11
c
21
/o
2
c
11
c
21
c
2
33
o
2
c
2
21
c
2
32
Goc
11
1/
2
oc
21
G
2
c
3
31
oc
21
G/
2
o
2
c
3
31
1c
2
11

o
2
c
11
c
21
Gc
3
31
o
2
c
21
c
2
33
Gc
31
oc
11
12/c
2
32
oc
11
1/c
2
33
oc
11
c
3
31
1Goc
11
1/Gc
3
31

o1/c
2
33
Gc
31
oc
21
G
2
c
3
31
/oc
21
`
5
`
6
Goc
21
`
5
`
7
`
6
Goc
21
1c
2
32
G/oc
21
`
7
`
5
G),(oc
11
1c
31

1Gc
31
/Gc
21
1c
31
G/ oc
11
c
21
Gc
21
),o
We can calculate c
2
31
, c
3
32
and c
3
33
xing the set `
1
= `
2
= `
3
= `
4
= 0 and sort
independently from uniform distributions (0.9: 0.9) the values of c
3
31
. c
2
32
. c
2
33
. `
5
.
23
`
6
and `
7
. Hence, each parameter of the matrices
1
,
2
and
3
are dened and we
can generate the DGPs of VAR(3) model with cointegration and WF restrictions.
24